Master Thesis - Department of Computer Science
Master Thesis - Department of Computer Science
Master Thesis - Department of Computer Science
You also want an ePaper? Increase the reach of your titles
YUMPU automatically turns print PDFs into web optimized ePapers that Google loves.
u<br />
2<br />
x 2<br />
u 1<br />
x 1<br />
(a) PCA basis (b) PCA reduction to 1D<br />
Figure 2.2: The concept <strong>of</strong> PCA. (a) Solid lines: The original basis; Dashed lines:<br />
The PCA basis; Geometric interpretation <strong>of</strong> principal eigenvectors illustrated in 2D<br />
space. (b) The projection (1D reconstruction) <strong>of</strong> the data using the first principal<br />
component.<br />
M is total number <strong>of</strong> pixels in the images and N is the total number <strong>of</strong> samples. Each<br />
<strong>of</strong> the face images xi belongs to one <strong>of</strong> the C classes {1, 2, ....., C}.<br />
• PCA (Principal Component Analysis): The key idea behind PCA [114,<br />
124] is to find the best set <strong>of</strong> projection directions in the sample space that<br />
maximizes total scatter across all images. This is accomplished by computing<br />
a set <strong>of</strong> eigenfaces from the eigenvectors <strong>of</strong> total scatter matrix St, defined as:<br />
x 2<br />
N�<br />
St = (xi − m)(xi − m)<br />
i=1<br />
T , (2.1)<br />
where m is the mean face <strong>of</strong> the sample set X. The geometric interpretation <strong>of</strong><br />
PCA is shown in Fig. 2.2. For dimensionality reduction, K (where K < M)<br />
eigenvectors U = [u1, u2, ..., uK] corresponding to first K largest eigenvalues<br />
<strong>of</strong> St are selected as eigenfaces. Reduced dimension training samples, Y =<br />
[y1, y2, ....., yN] can be obtained by the transformation Y = U T X. Now, when<br />
a probe image xt is presented for identification/verification, it is projected on<br />
U to obtain a reduced vector yt = U T xt. A response vector <strong>of</strong> length C,<br />
R(xt) = [r1, r2, . . . , rC] is calculated by measuring distances from the probe to<br />
the nearest training samples from each class. The distance function between<br />
13<br />
u 1<br />
x 1